[Qemu-devel] Re: AHCI support integration

[Tejun Heo]

Hello,

On 05/09/2010 09:11 PM, Alexander Graf wrote:
> Sebastian Herbszt wrote:
>> The ICH6 AHCI implementation submitted by Chong is an all-in-one
>> attempt (ahci.c).  It includes all needed parts of the ICH6, AHCI,
>> SATA and ATA specification.  The code in hw/ide/* on the other hand
>> is split (or could be split) into smaller parts like port based and
>> bus master access, IDE and ATA core.  I think it might be
>> reasonable to split ahci.c into ICH6, AHCI and SATA parts and strip
>> the limited ATA support and reuse it from the ide core.  This would
>> give us something like the following:
>>
>> hw/ide/piix.c (PIIX3 and PIIX4)
>> hw/ide/pci.c and core.c (IDE, BM)
>> hw/ata-core.c (ATA)
>> hw/sata/ich6.c (ICH6)
>> hw/sata-core.c (SATA)
>> hw/ahci-core.c (AHCI)
>>
>> Should this be a goal or am i over-engineering here?
> 
> [CC'ing Tejun - he volunteered to help out on this topic as well]
> 
> I think there's no need to split sata and ahci.
> Apart from that, I think we should take things incrementally. For now
> there's no need to split IDE further until we hit a technical limit. I
> have yet to see a patch trying to reuse the IDE command processing, so
> depending on how the respective person implements that, I'm open to
> suggenstions.

I don't know the code but here are a few things which might worth
considering.

* ahci and the IDE interfaces are pretty much independent in most
  implementations.

* On host emulation side, there might not be too much to separate out
  for generic sata part from ahci at this point.  Link state emulation
  should be pretty simple and I suppose multiple command processing
  would be a bit strongly tied into specific host emulation.

* If necessary, a separate IDE layer below PIIX3/4, ICH whatever would
  be nice.  Most IDE controllers behave about the same after all and
  depending on implementation except for link emulation and NCQ
  handling, ahci can probably just wrap the IDE layer for other stuff
  too.

* If necessary, implementing generic ahci would be nice.  Most ahci
  implementations only have small quirks on top of generic ahci
  behavior (even less deviance compared to IDE), so separating it out
  shouldn't be difficult at all.

Thanks.

-- 
tejun